Harvey has spent months arguing that law firms should not have to rent frontier intelligence from whoever happens to own it. This week it offered a concrete answer: a research preview of its own model, built on open weights, tuned specifically for the long, multi-step work lawyers actually do.
The model, called Tenet, starts from Kimi K3 and was trained with Fireworks. Harvey applied asynchronous reinforcement learning aimed at long-horizon legal tasks, blending synthetic content, public legal documents, and expert input while leaving customer data out of the mix. On the company’s own benchmark, Tenet roughly doubled the number of completed held-out tasks versus the base model and added 20 points on the contracts track, with all-pass rates up 9 and 2 points respectively. Harvey positions it as the current leader on the contracts portion and second overall.
The encouraging part is generalization. Tenet lifted results on Mercor’s APEX Agents and Crosby’s Redline Bench even though neither appeared in training, which suggests the tuning did not just memorize the test. Harvey’s broader argument is that open-weight models can match proprietary ones for specialized legal work without locking firms into a vendor’s walled garden.
For now, though, the company is sharing the method rather than the model itself. No weights, model card, or API endpoint have been published, and Harvey says production rollout inside its products will come later.
